Recursive Self-Improvement and the Future of AI

Summary

• The founders of an RSI startup—Dhaval Adjodah, who did a PhD in reinforcement learning at MIT Media Lab, and Owen He, formerly a DeepMind research scientist—say they are seeing empirical glimmers of recursive self-improvement: performance improving while the resource denominator stays fixed or falls. Asked whether this is happening, Owen says, “I don’t think it’s just possible. I think it’s happening”; they say results have grown over the last few months and they have held back releases until they are sure the results are correct.
• Their definitional test distinguishes RSI from “AI creating AI”: ask what the optimizer and the optimizee are. Humans using better AI tools is “autocatalysis”—the ideas and improvement still come from human brains. RSI, they argue, would put the language model itself in the optimizee role, closing the loop; in their framing, marginal gains could then stay constant or grow.
• Casado frames capability improvement in terms of performance over cost, but a closed loop alone doesn’t establish takeoff. A fixed numerator leads to a plateau; if it grows with capability, improvement can be constant, and if it grows as a higher power, the model implies exponential takeoff. A founder says the relevant exponent must be measured empirically; they lack a learning theory of deep learning.
• The denominator matters when evaluating RSI claims. Casado notes that model release times are falling, but says the trend looks different relative to capital raised. The founders track experiment costs—including human effort—on a “bleederboard.” They cite nanoGPT: GPT-2, once taking weeks on multiple GPUs, can now be retrained in hours on the same GPU for about $100, using the same data. Casado’s rejoinder is that he has yet to see the equivalent for GPT-5; the founders add that efficiency can buy more rollouts and better training trajectories.
• Their commercial wedge is inference efficiency, an AI problem they believe can improve both performance and cost. Faster inference also lets a system do more inference; the founders say their multi-LLM agent can choose what to optimize, and even its master prompt has been improving. They aim to optimize across prompts, harnesses, weights and chips, rather than isolated components.
• They say many published RSI ideas don’t scale in their tests, and practical bounds remain unclear. Their framework includes human researchers as agents and is intended to generalize across problem domains, though physical processes and the slowest component constrain progress. They contrast fast-evaluation areas like coding and math with drug discovery’s wet-lab and clinical-trial bottlenecks. Owen speculates that the universe’s computation sets an ultimate ceiling; he says he has yet to see a task-specific bound on the energy needed to solve a problem.

Deep dive

1. RSI is not “AI building AI” — it’s the optimizer becoming the optimizee

• Their definitional test for pitches Casado hears, such as “we use AI to create GPU kernels,” is to ask “what is the optimizer and what is the optimizee.” Humans using ever-better AI tools is “autocatalysis”: the ideas and improvement still come from human brains. In their view, RSI puts the language model itself in the optimizee role, reinvesting cognitive improvement into itself.
• Casado points out that every technical epoch had autocatalytic effects—computers helping build computers, for example. The founder’s distinction, via control theory, is that autocatalysis still needs humans to fill a gap, while a closed loop could make marginal gains constant or larger over time.

2. The capability-improvement model separates plateau, constant improvement and takeoff

• Casado frames the rate of capability improvement in terms of performance (the numerator) over cost or difficulty (the denominator). With a fixed numerator, as in his steam-engine example, the system plateaus. If the numerator correlates with capability, the result can be constant improvement; if it correlates with a higher power of capability, the model yields exponential takeoff.
• Casado cautions that a loop diagram alone doesn’t establish takeoff. A founder says it is hard to know what a nonlinear system will do and that the exponent must be measured empirically. They lack a good theory of deep learning; scaling laws are among the closest available guides, and are empirical too.

3. The denominator matters — count experiment costs and human effort

• Casado observes that model release times are decreasing, but says that trend no longer holds when considered relative to the capital companies have raised. The founders say improvement that requires pouring in more resources is not RSI, and that they include human effort and experimentation in the denominator. They track experiment costs on an internal “bleederboard.”
• They cite nanoGPT: GPT-2, which originally took weeks to train on multiple GPUs, can now be retrained in hours on the same GPU hardware, with the same data, for about $100. The improvements include algorithmic, data and kernel techniques; many are auto-research results. Casado’s rejoinder is that this is not yet demonstrated for GPT-5. The founders add that improving efficiency can enable more rollouts, better trajectories and smarter models.

4. A unified framework and the full stack matter

• The founders say many papers claiming RSI fail when tested in their internal infrastructure at scale on benchmarks and real economic problems. They also say that results announced in the field can prove to be noise or cherry-picking when they try to replicate them. Their view is that the simpler system that closes the loop is what scales.
• Casado invokes Feynman’s distinction between “stamp collecting and physics.” A founder argues that without a unified mathematical framework, researchers can’t tell whether the overall system is improving or only one component; optimizing one part may affect another.
• RSI can span prompts, harnesses, weights and chips. The founders started with harness improvements, then added weight updates, and expect to co-optimize multiple parts. Their agent chooses what to work on, and even the master prompt has been improving.
• They chose inference efficiency because it is close to AI itself and can improve both performance and cost. Faster inference means the system can do more inference; the founders also describe efficiency gains as enabling more rollouts and better training trajectories. Serving the latest open-weight models, such as Kimi K3, is also an economically important problem.

5. “It’s not a belief — it’s empirical measurement”: glimmers, variance and limits

• Asked whether performance can keep rising while the denominator stays fixed or falls, Owen says, “I don’t think it’s just possible. I think it’s happening.” He says they have seen results grow over the last few months, but have not released results until they are sure they are correct. He also notes that benchmark accuracy is capped at 100%, so it may not be the right metric for detecting exponential progress.
• Their measurement culture includes an engineer nicknamed the “Viceroy of Variance,” whose job is to try to break their results. Owen says there is no substitute for “quality time”: running hundreds of experiments to distinguish real effects from variance and reward hacking, including cases where an agent two levels down has hacked the process. Dhaval describes stochastic noise, oscillations and even different behavior between two B200 GPUs.
• Owen says intelligence is not a single scalar but a multidimensional vector: some areas may take off while others plateau, with progress constrained by the slowest component. Dhaval applies Amdahl’s law: speeding up digital work by a thousandfold or a millionfold has limited impact if it is only part of the total process. Drug discovery remains bottlenecked by wet-lab work and clinical trials, unlike coding and math, where evaluations are faster and cheaper.
• Owen says he has yet to see a quantitative bound on the energy needed to improve a system by a given amount or solve a given problem. He speculates that physical limits ultimately bound computation, citing roughly 10⁸⁰ particles and the speed of light as inputs to an upper bound. Dhaval’s “garden hose company in Wisconsin” joke suggests that important RSI applications may be hidden in unfashionable workflows; he compares this to engineers treating relativistic effects as an empirical correction before understanding them as physics. When Casado asks when to expect more, Dhaval says, “Very soon.”